On three training machines with the same GPU class, the bits per byte of Qwen2.5-0.5B and of its APT4 FVT transplant before recovery pretraining on ten evaluation texts are identical, and so are both models' digit-probe outputs.

Relation: Identical across machines

models

Qwen2.5-0.5B; its APT4 FVT transplant before recovery pretrainingModels scored.

machines

3 machinesMachines compared.

values identical

20 valuesBits per byte values identical on all machines.

values compared

20 valuesBits per byte values compared.

output files identical

2 filesPer-item digit-probe output files identical on all machines.

output files compared

2 filesPer-item digit-probe output files compared.

Method and evaluation protocol

Equality of the time-zero rows of the three arms' score files and of the sha256 of their per-item digit-probe files.

Dataset

The first 200 documents of ten evaluation texts: Polish and English web holdouts, English arXiv abstracts, LaTeX method sections, Python code and GSM8K problems, Polish Wikipedia science articles, PES questions and reviews, math_clean.Version: unspecified · Access: restricted

Reported results

20 of 20 values and 2 of 2 files identical.

Uncertainty and replication

Exact comparison.

Evidence references

bpb_mix00.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/bpb_mix00.jsonlbpb_mix10.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/bpb_mix10.jsonlbpb_mix30.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/bpb_mix30.jsonlmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/metrics.json

Limitations

Identical machines of one GPU model; a different accelerator would not reproduce the values bit for bit. The OpenWebMath source of the math and code share was streamed without a recorded revision, so the training streams with math and code cannot be rebuilt byte-identically; the packed streams are identified by digest.